Retrieval-Augmented Generation and DeepSeek-R1 Large Language Models for Implementing Financial Report Chatbot for Companies Listed on the Indonesia Stock Exchange

Ivana Lucia Kharisma (Informatics Engineering Study Program, Nusa Putra University, Indonesia)
Haldies Gerhardien Pasya (Informatics Engineering Study Program, Nusa Putra University, Indonesia)
Alun Sujjada (Informatics Engineering Study Program, Nusa Putra University, Indonesia)

Jurnal Teknik Informatika (JUTIF)
Vol. 7 No. 4 (2026): JUTIF Volume 7, Number 4, August 2026

The Indonesia Stock Exchange (IDX) requires all listed companies to publish financial statements regularly to ensure transparency for investors and stakeholders. However, the increasing volume, complexity, and technical language of these reports create significant barriers for timely analysis, particularly for retail investors, regulators, and academic researchers. This limitation highlights the urgent need for intelligent systems capable of automating the extraction and interpretation of key financial information. This study aims to design and evaluate a Retrieval Augmented Generation (RAG) model to process financial reports, specifically the audited 2024 financial statements of Bank BCA and Bank Mandiri, in order to develop more efficient AI-based tools for managing complex financial documents. The research adopts a quantitative experimental approach using income statements from PT Bank Central Asia Tbk and PT Bank Mandiri (Persero) Tbk for the years 2023–2024. The methodology involves multimodal text extraction with Gemini Flash 2.0, preprocessing and cleaning using Regular Expressions (Regex), document chunking, and embedding generation with the multilingual-e5-small model. A Qdrant vector database is used for storage and retrieval, while DeepSeek-R1 serves as the transformer-based LLM for generating responses. Model performance was evaluated using BERTScore and ROUGE, supported by expert assessment. The system produced strong results, with BERTScore averages of 0.8074 (precision), 0.8253 (recall), and 0.8156 (F1-score). ROUGE evaluation yielded average scores of ROUGE-1: 0.5986 and ROUGE-2: 0.4531. Expert evaluation confirmed high parsing accuracy (94.6%) and the system’s ability to generate accurate, contextually relevant, and multilingual responses. Beyond its practical application, this study contributes to the advancement of scientific knowledge by providing an integrated framework for multimodal financial document processing using RAG-based LLMs. Overall, the study successfully developed a RAG-based LLM chatbot capable of effectively extracting, processing, and answering queries related to financial reports, demonstrating strong semantic alignment and supporting the development of more accessible AI-driven financial analysis tools.